Upload
vandat
View
214
Download
0
Embed Size (px)
Citation preview
1
KNOWLEDGE SEMANTICS
Preface
Linguistics emerged as a science in the mid 1950’s when Noam Chomsky, a
young professor in the MIT Research Laboratory of Electronics, gave a series of
lectures which culminated in his book: Syntactic Structures (1957). In the then
broad field of linguistics, which mainly studied bible translations, language
variation, the Indo-European and other language families, historical linguistics,
animal communication, bee languages, and sundry other topics, his book, actually
a small soft-cover monograph of about 100 pages, fell almost stillborn from the
press. A brave publisher, Mouton, printed several hundred copies, and Chomsky’s
contract required him to buy all the unsold copies after a year or two.
In the MIT Research Lab of Electronics, his book caused reverberations,
which, to get ahead of our story, resulted in Syntactic Structures selling tens of
thousands of copies and being translated into dozens of languages. How and why
did his book electrify the nascent fields of what became computational linguistics,
cognitive psychology, signal processing theory, and computer science? George
Miller, a close friend and colleague of Chomsky, presents a readable version of the
early history and impact on psychology and education (“The Cognitive
Revolution,” TRENDS in Cognitive Sciences Vol.7 No.3 March 2003). When I first
read Syntactic Structures (1963) I was working on a PhD in electrical engineering
to develop a ‘computational typewriter’ such that one would speak into it and it
would type your text. After reading the book, I told my advisor that the device
could never be built using the incorrect (finite state grammar) model we were
using. I simultaneously applied to the MIT Linguistics Department to work with
2
Chomsky. I joined the MIT Linguistics Department, graduating in 1968 with a
PhD advised by Chomsky.
In 1964 I earned a Master of Science degree at Dartmouth College Thayer
School of Engineering where we used the GE 235 computer, probably one of the
biggest in the world at that time, to solve problems in vibrating systems, and in
particular, how to dampen vibrations in battleships when the big guns went off, the
shaking in airplanes during turbulence, the steering of submarines and torpedoes,
and so on. General Electric, Dartmouth, and MIT (as well as others) teamed up to
develop the first ‘time-sharing computer’. It had 60 or so remote terminals linked
to it by coaxial cables, and а person a mile or so away could do all computations
remotely and have the results transmitted back to their local printer. Prof. John
Kemeny headed the project. The early history of the 235 time-sharing project led
to interdisciplinary work between almost all departments of all the universities
involved. Circuit design and power supplies dominated my engineering work, and
early on they had me doing the ‘interfacing’ of keyboards, printers, and plug-
boards. The operating system of the GE 235 was called multics, or MULTIX. After
the project was killed, two young men took the basic idea to Bell Labs and called
their project UNIX.
Early computers did have ‘languages’, but much of the ‘programming’
constituted itself as ‘running wires from here to there’, or plugging cables into
holes. Look up the German Enigma Code machine and you can see what type of
‘cutting edge’ devices we thought about. Some early thinkers, Alan Turing, John
von Neumann, Claude Shannon, and Norbert Wiener, all of whom had an
enormous influence on me and the groups I worked with, thought very little (very,
very little) about ‘hardware’ such as ‘wires’, ‘switches’, and ‘plug-boards’. They
3
talked about ‘computer languages’, which at the time seemed an arcane concept.
Alan Turing, often called the ‘father of modern computing’, developed a computer
(often called ‘computational engine’) during WWII that cracked the German
Enigma Code machine that secretly directed the operations of the Luftwaffe and
the German submarine fleet.
In September 1940, England, fearing a German invasion, decided to send to
the USA all of its military secrets, secret weapon plans, and researchers. The
readable article, Groupthink (Jonah Lehrer, The New Yorker, January 30, 2012),
lays this out in some detail, and indicates that many, even most, of the English
scientists and engineers took up residence at MIT and did their research in MIT
Building 20. Groupthink shows how this group defined the Research Laboratory of
Electronics, and details how Chomsky’s Linguistic Studies grew out of the rarified,
heady air of the forms of computation and symbolisms that gave rise to radar,
sonar, artificial intelligence, and expert systems. In short, Chomsky offered a
theory, called Generative Grammar, which provided a conceptual framework,
detailed notation, and methodology to take all of the formal systems thinking and
image processing ideas from radar and sonar and express them in a way that
offered a clear exposition for some aspects of human language thinking and
language use.
In this era (1950-1960, 61…) no clear concept existed of ‘computer
language’ or how to program a computer to do anything. The ideas of Alan Turing
(illustrated well in the movie, The Imitation Game, starring Benedict Cumberbatch
and Keira Knightley) came to dominate all thinking over the past decades. Turing
wrote little, and what he wrote, aims to educate experts. No sugar coating exists
anywhere. John von Neumann, one of the world’s most brilliant mathematicians,
4
offers a view of Turing’s work that relates directly to the ideas of Chomsky, and
probably the work of just about everyone else that mentions Turing and computers.
People may argue, but the current ideas of ‘computer’ and ‘computer language’
derive almost totally from John von Neumann’s writing, which range from the
popular (The Human Brain) to the outlandishly complex (Theory of Self-
Replicating Machines). The best introduction to John von Neumann exists in his
voluminous publications and the writings of his ubiquitous students. The best
introduction to Turing for beginners exists in von Neumann’s writing about what
he thought Turing accomplished.
One could say that Chomsky’s views of generative grammar and Syntactic
Structures reflect John von Neumann’s views of Turing’s research. Let us just take
one idea discussed at length by von Neumann: Any computer of any interest
always has two different languages, and, the main function of the computer is to
translate sentences from one language into the other. (I believe this is the best way
of thinking about Turing.) The two languages may use the same symbols {a, b,
c…, space, period, question…}, but each language defines a different set of
sentences. A machine that translates from English to German, or vice versa, fits
this system. The German Enigma encoding/encryption machine input a ‘normal’
German sentence, scrambled the letters, and produced an encoded sentence of
apparently random letters. When Turing entered the encrypted sentence into his
Turing machine, the output constituted the original unencrypted sentence.
In Turing’s system, the only thing a computer does is convert one string of
symbols written in alphabet X into another set of symbols written in alphabet Y,
where X and Y may or may not be the same alphabet. Here is an ‘in-joke’ among
Turing aficionados: The German Enigma machine converted a ‘plain-text’
5
German sentence into a random string of letters from the German alphabet; but,
and this is a giant ‘but’, in the encoded text, no letter could be encoded at the
original letter in the plain-text. So, if the input contained the letter ‘a’, the random
output encryption could not have an ‘a’ to represent the encrypted ‘a’. This was an
Achilles Heel enabling Turing to ‘break into the Enigma’ system. Turing said: A is
not A to the Enigma. This is a contradiction. From a contradiction we can derive
any conclusion. This was Turing’s way of saying that the Enigma had a flaw, a
keyhole, an Achilles heel. I thought it was funny when I heard it. Turing made
great use of this fact, but his argumentation tends towards the difficult.
What came to be called cognitive studies took over this perspective and
assumed the Turing/von Neumann computer could model some aspects of the
human mental capacities. In many, perhaps most, studies, the Chomsky
interpretation of the Turing/von Neumann computer view dominated. Chomsky
argued that the grammar of a human language defined the speaker’s knowledge of
the language. The grammar, being realized as neurons and assorted brain input-
output mechanisms (ears, hands, mouth, tongue…) defined a computer. Hence,
following our discussion, the human brain/computer might in fact have two
different formal computer languages defining its operation.
In this discussion, the term language (a technical term with a specific
definition: ‘a sequence of elements’) includes all what are called Natural
Languages (such as English, French….), and often these are called Languages,
with a capital ‘L’. But language also includes the language of telephone numbers
(there are valid phone numbers and invalid ones). It includes all chemical
formulae, all computer languages, and so forth.
6
The term Language (English, Chinese…) actually decomposes into many
different languages (lower case ‘l’), and these ‘languages’ represent specific
capacities of a normal speaker of a language. Since Syntactic Structures first
appeared, the use of terminology has become under control, and instead of talking
about ‘many languages’, most people say ‘components of the mind’: a sound
component, a syntactic component, a meaning component, and so on. Internal to
each component is some formal system of representation, and this formal system of
representation is a ‘language’. Look in the beginning of a large dictionary and
there will be a ‘pronunciation guide’ using about 150 symbols, each defining a
specific sound. Then each word in the dictionary has these symbols in a string to
show you how to pronounce with word you look up. The pronunciation guide
defines a language similar to that in the phonetic component of a Chomsky
grammar.
A school for actors, translators, or opera singers may teach only spoken
Italian, German, Russian, and so on for people that do not care about meanings or
even making simple sentences. They simply want to sing the opera with perfect
pronunciation. Phonetics studies the ‘language’ of basic sounds and how they go
together. Phonetics defines the study of the sounds of a human language,
completely independent of any meanings or interpretations. Speech therapists,
opera trainers, and others study and teach phonetics. The Professor in ‘My Fair
Lady’ was a phonetician.
Descartes and Pascal back in 1660 argued about whether or not one should
move the lips and make sounds while praying. Not a big deal to most people today,
but in 1639, when the Catholic Church declared the works of St. Augustine, the
founder of the Catholic Church, to be heresy, the wrong answer could get you
7
killed. Descartes said, no lip sync, the ideas going through your head led to
salvation and blessedness, not the sounds you make. Pascal in his Pensées
denounced Descartes as a heathen for his concepts of silent prayer. Their mentor,
St. Augustine, gave; sermons indicating that one might start a prayer in one
language, switch to another, and then end it in a third. The flow and sequence of
ideas coursing through the mind led to salvation, not the particular sounds or
mouth movements. Semantics has traditionally been the study of the meanings
expressed by languages, and in a particular language, independent of the sounds or
sound sequences in the language. So, in English, Sue wondered whether or not to
go and Sue wondered whether to go or not have the same meaning (are
synonymous), but differ in the sequence of sounds. Going back centuries
(Descartes) and millennia (St. Augustine) Chomsky’s ideas would agree with those
that said the ‘sounds’ of the prayers were divorced from the ‘meaning’, or ‘healing
power’, or ‘salvation power’ of the prayer. Many religions are not like this: You
must exactly pronounce the prayers as they did hundreds of years ago.
According to Chomsky, each human Language has a Grammar, but this can
be factored into parts. For a Language with its Grammar, there exists a theory of
phonetics (a sub-grammar of phonetics) and a theory of semantics (a sub-grammar
of semantics). So, following von Neumann, a human Language has a mental
representation in a brain as a Grammar, but the Grammar consists of two sub
grammars (often called components), The phonetics component defines the sounds
of the Language; and this relates to the (sensor-motor) parts of the brain that
control the ears, mouth, and hands (sign language). The semantics component
defines the conceptual correlations, or sematic representations, of the Language.
How the semantic information of Language exists in the brain has always been a
mystery that has spanned disciplines. However, almost all agree that the words and
8
sentences of a Language point to things, actions, and events in the world in some
ways.
How does the phonetics component relate to the semantic component? That
is, how does a Grammar link specific sound sequences (like The dog bit Sue) to
specific meanings that we can judge to be true or false about events in the world?
Dr. Tarnawsky focuses on this problem: How do the sequences of sounds
(or written letters and words) relate to meanings, and these meanings relate to
observable objects and events in the world?
When he wrote his dissertation, almost everyone in the world thought that
each sentence of a Language had a phonetic representation (like the pronunciation
given in a dictionary using special symbols) and also a semantic representation,
which would be some graphic representation that related directly to some explicitly
definable conceptual scheme. Unfortunately, as Dr. Tarnawsky points out, no one
ever offered any type of plausible scheme for representing meanings in any
language for any wide range of concepts. Some progress had been made for
representing some aspects of meaning using formal logic and various theories of
quantification using numbers. These offered insights, but the representations did
not seem to be ‘representations of meaning’ in any broad sense. They were simply
a shorthand representation in no specific formalism of any generality.
In 1982, usually examples presented in the literature pointed up problems
with ‘meaning’ inconsistencies in a human language, and then tried to show that
some specific notational convention for semantic representations could show or
clarify the anomaly. Most examples did not show any ‘meaning’ in the abstract,
9
but offered ways to understand why some sentences were strange or did not have
the meanings they should have. Consider an example.
Suppose there is one telephone, but two people involved: Tess and Tracy.
One can say: Either Tess or Tracy does not have a telephone. But one cannot say:
*Either of them does not have a telephone. One can say: Tess does not have a
telephone, but Tracy does. Or, Either of them can use the telephone. Why is the
one sentence odd? It is perfectly good English and has a phonetic representation.
But semantically, somehow, the logic does not work out. This is a failure in the
semantic component.
Suppose there is a room containing a lot of people and I look in. I might say:
Some of the people are taller than others. Or, Some of the people are shorter than
others. But it is odd to say: *Most of the people are taller than others, or *Most of
the people are shorter than others. Both of them are true. If there are 20 people of
different heights, then 15 are either taller or shorter than the other 5.
Much work in ‘formal semantic studies examines such constructions (which
occur with similar problems in all languages). Usually the researcher offers some
method/procedure to convert the sentence into some formal logical notations and
try to show the formal notations lead to contradictions. Dr. Tarnawsky argues that
while some such studies may produce insights in some narrow areas of language
(those using logical words like each, all, some, most, not…), the work offers no
general concept of semantics to apply to nouns, verbs, and general vocabulary.
A lot of research focused on types of ambiguities. For instance, the string of
words: John decided on the boat contains an ambiguity: Either John decided to
10
choose the boat or John, on the boat, made his decision. Chomsky’s theory offers
two syntactic structures for the one sentence: John (decided on) the boat, with the
complex verb decide on as a unit; and also, John decided (on the boat), with on the
boat a description of location. The semantic difference arises from the different
syntactic structures the grammar assigns to a single string.
So far, almost everyone working up till 1957, when Chomsky published his
work, would pretty much agree with the above general outline. Chomsky argued
that in addition to a phonetic representation (the sound) and a semantic
representation (the meaning), each sentence had a syntactic structure, which
indicated how the phonetic representation was broken into individual
sounds/letters, how they were syntactically grouped into words, how the words
grouped into phrases, and these phrases into sentences. To make a long story short,
Chomsky showed rather conclusively that precisely the computational mechanisms
and notations used by Alan Turing, John von Neumann, and Claude Shannon were
ideally suited to define the syntactic structures of any human language.
Chomsky persuasively argued that each sentences of every human language
could best be thought of as having a graphic representation (formally, a branching
network structure in computer theory) and if we knew this ‘abstract’ structure, we
could read off of it the phonetics (sound) and semantics (meaning) of a sentence.
Traditional linguists and semanticists criticized him mercilessly. The most
interesting and witty critique came from an English literature professor with the
title Syntactic Strictures, mocking Chomsky’s book title Syntactic Structures.
Chomsky won the day, and today a human language is considered to be a set of
syntactic structures, each of which corresponds to a sentence, and it gives all
11
properties of the sentence relating to sound, pronunciation, meaning, and pragmatic
usage.
Dr. Tarnawsky’s work offered crucial evidence to support the overall view
that syntactic structures, these actually rather abstract mathematical and
computation objects, were ‘real’ and not simply a convenient fiction. How did he
do this? By showing that the level of semantic representation could be done away
with.
While most (perhaps all, including me) people believed that each sentence
was divided into three different components (phonetic representation, semantic
representation, and syntactic structure), Dr. Tarnawsky advanced the view that we
could abandon the level of semantic representation and answer any and all
questions (that existed at the time) being asked, using only the syntactic structure
and the phonetic representation. This not merely offered a ‘simplification’ of the
theory, but gave a profound insight into the organization of human language.
While meanings play a crucial role in human language design, they are not to be
considered as a part of the human Grammar. They arise when the human
brain/mind pragmatically applies the sentences (syntactic structures) to
observations looking for appropriateness and usefulness of the sentence in specific
situations and contexts.
Dr. Tarnawsky offered conclusive arguments in his dissertation, but they
tend to be subtle and involved. They do not hinge on properties of his notations, or
those that existed in the 1980s, and are as valid today as then.
12
Dr. Tarnawsky’s work possessed implications beyond the narrow range of
examples presented in the literature of the time. His dissertation helped cause a
shift in the balance of power among research departments. Basically Noam
Chomsky attempted to make the ‘underlying formalisms’ of linguistic theory
derive from the computational theories of Turing, von Neumann, and Shannon, all
workers in the MIT Research Laboratory of Electronics, and all workers on war
projects such as radar, sonar, submarine detection, and code cracking. The bulk of
the researchers Dr. Tarnawsky argued against derived their ‘underlying
formalisms’ from 19th century philosophers and logicians. Those that advanced
‘semantic theories’ in opposition to Dr. Tarnawsky’s tended not to come from
‘centers of computation’ but from philosophy departments, Literature departments,
and clinical psychology research centers. Most of the opposition was from
‘traditional psychology’ and classical philosophical views. Dr. Tarnawsky’s
formalisms were firmly in the computational views of symbolic algebras of the
Research Laboratory of Electronics.
Knowledge Semantics offers a clear view of Dr. Tarnawsky’s basic thesis,
and his examples, drawn from several languages, illustrate his ideas perfectly. Let
us examine some of the problems clarified by the dissertation.
Pronoun co-reference (anaphora) defined main problem in extracting
meanings from the syntactic structures. Very often, two almost identical sentences
not only had different meanings, but the references of the pronouns became
convoluted. Consider these sentences:
(1) John appeared to Mary to perjure himself.
(2) John appealed to Mary to perjure herself.
13
(3) *John appeared to Mary to perjure herself.
(4) *John appealed to Mary to perjure himself.
In these next sentences, which are equivalents of the above, with Bill and himself
replacing, respectively, Mary and herself, italics mean that the words refer to the
same person; they are co-referential:
(5) John appeared to Bill to perjure himself.
(6) John appealed to Bill to perjure himself.
(7) John appeared to Bill to perjure himself.
(8) John appealed to Bill to perjure himself.
Dr. Tarnawsky did not only offer a new way of thinking about such
problems, and a new type of notation for expressing them, but he offered novel
argument structures and types of arguments for justifying any analysis that
attempted to ‘unify’ a view of syntax and semantics.
For instance, when he wrote, almost all linguists would contend that
sentences 1-4 posed problems in noun-antecedent agreement of
masculine/feminine pronouns, and that the problem was one of syntax, not
semantics. But for Dr. Tarnawsky, the problems posed by 5-8 are identical to those
posed by 1-4, and also have a syntactic solution. The distribution of data is
described by syntactic mechanisms of agreement with no concern for meanings.
Those with a level of semantic representation argued that while 1-4 were
syntactic agreement, 5 -8 were pronoun referent problems and the antecedent of a
pronoun was determined in the semantic component. But clearly, the syntactic
mechanisms can treat 1-8, and other sentences like, We appeared to you to perjure
14
ourselves, *We appealed to you to perjure ourselves, and so on as simple syntactic
agreement problems.
No motivation exists to set up a whole new level of semantic representation
just to handle cases where the pronouns happen to be of the same person, number,
and gender, such that we can label this a problem of ‘co-reference’. Syntax defines
‘antecedent’ and ‘referent’ and this does everything in all cases. In English
syntactic and semantic gender tend to be the same. But in some language, like
German, one has das Mädchen (the girl) which licenses the pronoun es (it) not sie
(her). Sentences involving such words must be handled by syntactic agreement
since the semantics is irrelevant.
The view of the syntax-semantic interface offered by Knowledge Semantics
goes beyond the insights obtained by studying simple paradigms of words and
sentences. Dr. Tarnawsky has always exhibited keen interest in problems of
translation and the ‘expressive limits’ imposed on us by our languages and
vocabularies. Chomsky argued that the main thrust of ‘Linguistic Research’ was to
explain the concept of ‘linguistic intuition’ both as it shows up in sentence
structures, but more profoundly, how it shows up in the nature of the overall
structure of the Grammar that defines and constrains the Language. Consider some
ways the constraints on vocabulary discussed in Knowledge Semantics helps us
understand difficulties in translations.
Mark Twain, during part of the 1890’s was the music critic of the New York
Herald Tribune. New York opera buffs were solidly pro-Italian, and Wagner had
been given short shrift in the tabloids. Twain, asked to review Wagner’s Ring,
opened his essay, Wagner’s music is not as bad as it sounds. This sentence has two
15
syntactic structures, and two meanings: it is not happy on the ears, and, not as bad
as people say. Twain’s sentence translates directly into German and preserves the
ambiguity: Wagners Musik ist nicht so schlimm wie es klingt. But it does not
translate into French as a single sentence with the same ambiguity. What does this
mean? How does this relate to intuitive grasp of a Language and Grammar?
How can one know it does not translate into French as a single sentence? A
non-native speaker might have to search for a word in the dictionary and
eventually give up. A fluently bilingual French-English speaker knows
immediately, via intuition, that it cannot have a single sentence translation. But a
fluently bilingual French-English linguist familiar with syntax and semantics can
assert that not only is there no existing translation, but there can never be given the
syntactic structures of French. There cannot be any new French word ‘invented’
that can support the syntactic structures simultaneously needed to preserve the
ambiguity in translation.
As Knowledge Semantics makes abundantly clear there is a fundamental
underlying architecture of the syntactic structures and another equally fundamental
architecture underlying the semantic pattering that exists in any language, and in
some cases, like those elaborated in the book, they correlate. But in general, the
correlations require only a level of syntactic structure to be defined, not any level
of semantic representations.
To step back and look broadly at Knowledge Semantics, it elaborates on
what is probably the main theme of Chomsky’s revolutionary Syntactic Structures:
the concept of a level. What we have called ‘components’ and ‘languages at those
components’ were called ‘levels’ in the opening lines of Syntactic Structures. It is
16
not that one is ‘higher’ or ‘lower’ in any sense, it is rather that they are abstract
ways of thinking about the same thing: Grammar and the Language it defines.
Consider maps describing the surface of the earth. There are at least three.
At the most basic level, we have the map of the giant plates of hardened lava that
float on the molten center core of the earth, the plate tectonic map. We also have
a geographic topological map showing mountains, valleys, rivers, and so on. The
third represents the political map, showing cities, states, nations, and lines like
meridians and equators imposed by human vanity on a system basically indifferent
to human values.
These three maps define abstract views of the very same object. Similarly,
the syntactic structure, the semantic/pragmatic views, and the phonetic view are
advanced as three views of a single object: a sentence.
It appears that the three different abstract maps of the earth are here to stay.
Each view treats a qualitatively different problem. Plate tectonics describes
earthquakes and volcanoes. The topological map described landslides and flooding
in tidal marshes as well as glaciers and ice at the poles. The political map accounts
for the locations of wars and human migrations.
But the three views of a sentence, which have dominated linguistics for
centuries, are not as certain. As Dr. Tarnawsky points out, no reason exists to have
a semantic level, since any task that was to be solved by the semantic
representations can be solved by the syntactic level using mechanisms
independently motivated, such as ‘agreement’. The semantic level is an
anachronism left over from earlier, pre-Turing computational, times.
17
Knowledge Semantics shows rather elegantly using persuasive arguments
and data from different languages that we lose nothing by abandoning the level of
semantic representations. Rather, we gain a more coherent view of the overall
structure and architecture of human language. And, if Chomsky’s mantra,
‘Language is the Mirror of the Mind’, is correct, then we gain a greater
understanding of the operation of the human brain.
Dr Tarnawsky’s thesis took very strong stands on many controversial issues,
hotly debated in the fields of linguistics, cognitive psychology, philosophy,
computational data base studies, and any discipline concerned with offering a
logical or mathematical model of human knowledge and language understanding.
He laid out the issues clearly, in the necessary detail, and treated all opposing
theories fairly and explicitly. The overall theoretical perspective would require
reading the whole dissertation. Here we touch on the types of issues raised, and
indicate the contribution of the thesis.
A basic focus of all Chomsky’s generative grammar was displacement, the
fact that in the semantic interpretation of a sentence, sometimes a word or phrase is
understood to be the subject or object of a verb quite far from it in the sentence. In
these sentences, John is the subject of please for eager, but the object of please for
easy. Sentence (10) means, it is easy to please John, not that John pleases. This is
an example of displacement, or often called ‘movement’ in 1982.
(9) John is eager to please.
(10) John is easy to please.
18
Some verbs, like ride, can take a direct object (a horse) or a prepositional
phrase object (on a horse):
(11) It is easy to ride a horse.
(12) It is easy to ride on a horse.
Sentences (12) and (13) illustrate another type of displacement. Notice that
they are synonymous, but a horse can be displaced from the on prepositional
phrase to the beginning of the sentence, and in fact, to the subject position, where a
horse replaces the pronoun it.
(13) A horse is easy to ride on.
Very strict rules—which everyone agrees are pure Chomsky-type syntax—
govern where displacement phenomena can occur. So, for instance, while what,
which is a noun phrase can ‘move’ to the front and replace the it, under no
circumstances can the prepositional phrase on what move and take the place of the
pronoun it.
(14) What is easy to ride on?
(15) What is it easy to ride on?
(16) On what is it easy to ride?
(17) *On what is easy to ride?
One of Dr. Tarnawsky’s fundamental contributions to linguistics as a science
arose from two insights. (1) If we abandon the concept of a separate level of
semantic representation and instead utilize the formal computationally defined
entity he calls a knowledge base, which today might be called a data base of
syntactic linguistic knowledge, then, we (2) can extend the concept of
19
displacement phenomena—well defined in syntax—to account for what had
traditionally been called ‘scope of quantifiers’ in semantic theory. The arguments
are developed systematically throughout the thesis, but especially in chapters 5 and
6 where he discusses the works of others.
Dr. Tarnawsky’s grammar which considers ‘scope of quantifier
interpretation’ to be defined by the same syntactic mechanisms called
‘displacement’, predicts that there is no sharp break between syntactic examples
and semantic examples. One can see this from examples like the following
involving the lexical item whether or not. The whether can be separated from the
or not by long stretches of words and phrases, usually by syntactic movement
rules. The same syntactic mechanisms that determine movement in examples like
the above, define the correct interpretation of these sentences with no need for any
level of semantic description.
(18) Mary cannot decide whether or not to go.
(19) Mary cannot decide whether to go or not.
(20) Mary cannot decide whether to go with the boy or not.
(21) Mary cannot decide whether to go to NY with the boy or not.
(22) Mary cannot decide whether … or not.
Where … can be long.
Examples like these involve quantifiers, since whether and or not are
quantifiers. It follows from Dr. Tarnawsky’s thesis and reasoning that there should
be a ‘continuum’ of problems of quantifier interpretation and scope, and they all
can be treated by the same syntactic, not semantic, mechanisms that define the
properties of displacement.
20
Sometimes the displacement is not overt, that is, no words seem to move or
be in the ‘wrong place’, but the interpretation shifts somehow from one part of the
sentence to another (the scope of quantifier problem). Consider these examples:
(23) How did you know to wash the car?
(24) How did you want to wash the car?
The answer to the first question could be Mary told me, but not with a
sponge. The answer to the second question could be with a sponge, but not, Mary
told me.
In the first sentence, the how (a quantifier) is understood to link with the
verb know, not the verb wash. In the second, how links with wash in the lower
clause and not the want in the first clause.
The syntactic mechanisms of Knowledge Semantics treat cases like these
where there is no overt movement or overt displacement by the same rules that
define the semantics of all cases discussed above.
Dr. Tarnawsky’s fundamental underlying thesis, which gives the title
Knowledge Semantics to his work, stems from his insight that the goal of sentence
interpretation is not to obtain some ‘level’ of formal notations called a semantic
reading, but instead to show that the mechanisms that actually relate a sentence to
the human’s knowledge base are independently defined syntactic operations. The
insight is that the meaning, or lack of meaning in some cases, derives from what
knowledge the speaker brings to the interpretation process. The focus is on the
21
‘processes’ of interpretation, not on some final ‘data structure’ that we might call
‘the interpretation’.
There are cases in which a sentence, which one might suspect should have a
‘meaning’, in fact does not seem to have the expected or in fact any meaning.
Those against Dr. Tarnawsky’s views would argue that such cases have an
anomalous semantic structure. Dr. Tarnawsky argues the rules that try to match the
sentence to the speaker’s knowledge base ‘jam up’ and fail to function. Examples
of such phenomena abound. Dr. Tarnawsky provides a perspective and tools to
deal with them.
The basic ideas of Knowledge Semantics defined a fruitful area for further
research into murky areas. So I will pose to you a question. Think for instance, of
what reasons can be given for the ill-formedness of *I don’t know why to go, given
that many other quantifiers form interpretable sentences. Having read and
understood Dr. Tarnawsky’s proposals, I suspect it has to do with trying to fit the
question into my knowledge base and not some quirk in some abstract logical
semantic structure.
(25) I don’t know how to go.
(26) I don’t know when to go.
(27) I don’t know whether or not to go.
(28) *I don’t know why to go.
(29) I don’t know why I should go.
PhD dissertations usually focus on one or two narrow issues that people in the
discipline believe can resolve a problem and put research into a channel that will
produce more insightful theories in the future. So far, we have focused on the trees,
22
specific mechanisms and data structures of ‘Linguistics’, but now let’s focus on the
forest: How does ‘Linguistics’ as a science relate to other sciences and broader
concerns in the philosophy of language? In doing so, we can see how Dr.
Tarnawsky’s ideas about scientific research, the goals of research, the methods of
research, and the ‘utility’ of research increase insight into the formal processes of
human language, which Chomsky linguists generally believe are ‘syntactic
processes’ which link sounds with meanings.
Suppose we ask a chemist, ‘What is sulfur?’ or ‘What is iron?’ She might
put a lump of yellow powder in front of you and call it ‘sulfur’, and place a cube of
metal on the table and label it ‘iron’. One might say these physical objects are the
‘representations’ or the ‘meanings’ of ‘sulfur’ and ‘iron’. Perhaps after this you
could decide if something is, or might be, ‘sulfur’ or ‘iron’. In general, this style of
thinking has been called ‘ostensive definition’. But this only scratches the surface.
Sulfur bears the properties yellow (eye), soft (touch), and smelly (nose) while iron
tends to be shiny, hard, and odorless. Certainly these properties of how ‘objects’
and ‘words’ relate to our senses must be part of the sounds and meanings here.
Also, people often build things out of iron, but few things out of sulfur. So part of
the sounds and meanings must have to do with ‘use’ and ‘function’. The meanings
of things and words do not seem absolute in some ways: Can one be a vegetarian
and eat fish, shrimp, and oysters? Is a crawfish ‘meat’ or ‘vegetable’? Probably
somewhere in between for some vegetarians. For those that eat anything on the
menu, the question makes no sense, and, insofar as it does, the answer has no
pragmatic significance since it will not affect behavior.
In the 1980’s, there existed two main viewpoints, usually hot button issues
in philosophy departments, and simply issues in other departments. In linguistics
23
departments following the ideas of Noam Chomsky, most linguistic research
assumed the ‘semantic representation’ viewpoint, then espoused by Katz, Postal,
Fodor, and Jackendoff, among most others. It was the prevalent view of MIT style
linguistics. The main alternative view in the 1980’s followed the ideas of Hilary
Putnam, a Harvard philosopher, who argued that semantic representations may or
may not, and probably did not, exist and were quite irrelevant to studying language
as a sound-meaning correlation. Everyone agreed that meanings exist, but one
group (Katz, et. al.) argued they have formal representations in some logical type
notation, the other group (Putnam) said they do not.
To use an analogy, Katz et al., somewhat like the chemist who placed two
objects on a table, argued that the formal objects they represented in their notations
were ‘real’ and constituted the ‘meaning’ of a sentence; Putnam is like the chemist
who would argue that the objects only tell a part of the tale. The objects indeed are
a representation, but only a part of the whole representation. We must also consider
how the objects appear to the senses, how they function in human context, as well
as where they come from, and many other things. For Putnam, there exists no
single thing (physical or as a formal abstraction) that we can point to and call it the
‘meaning’ of a sentence.
Putnam, an excellent writer, said that he did not know the difference
between an elm and a beech tree, but he could use the words ‘meaningfully’ to
differentiate two ill-defined concepts. His theory had a distinct ‘social’ aspect:
Putnam said he could use the terms confidently because there were ‘experts’ in the
English-speaking world that did know the difference. It made a difference in their
usage of the words, but not in his usage.
24
Dr. Tarnawsky’s dissertation, with its strong emphasis on the ‘social and
contextual usage’ of sentences falls squarely into the camp of Putnam, who was
rarely discussed by Chomsky MIT linguists. Dr. Tarnawsky’s dissertation ran
counter to 99 percent of semantic research in the Chomsky MIT school of
linguistics. It was revolutionary in offering a novel perspective and a detailed
computational theory that merged Chomsky’s and Putnam’s views.
To use an analogy, one of Dr. Tarnawsky’s criticisms of the ‘semantic
representation’ group argued that all they did was ‘translate’ from one notation into
another, but gave no insights into what a ‘meaning’ could be. Consider numbers
and how they might be represented in various notations. Alan Turing would count
using symbols like these: { } Romans used
these: {I, II, III, IV, V, VI, }. Fibonacci introduced the Arabic numerals which we
use today: {1, 2, 3, 4, 5, 6, ). With the advent of computers, binary numbers
achieved popularity: {1, 10, 11, 100, 101, 110, 111, }. Which notation is ‘true’ or
‘correct’? The question makes no sense since they all stand for the same things—
numbers.
A single young lady named Jane Bloom marries John Doe and instantly
becomes Jane Doe for legal and economic purposes, such as taxes, health
insurance, and the right to stay in their rent controlled apartment if John Doe, the
leaseholder dies. These label changes simply follow from the translation laws of
the marriage legal code. We have not learned anything about the lady represented
and probably, except that she may now have a ring on her finger, she remains
otherwise unchanged. We know that Señora Jane Doe qualifies for a green card
and USA citizenship, but Señorita Jane Bloom does not.
25
Reflecting on the understanding of the concept ‘number’ in the different
representational notations above, we know ‘numbers’ no better than before, but we
now have different ways to write down numbers. Each ‘representation’ has its
advantages for specific reasons, and pragmatically one may surpass the others. But
these representations are simply ‘notational variants’ of each other. If the Chinese
and Japanese abandon their graphic scripts and use the Cyrillic or Roman alphabets
to represent words in their languages, some simplicity may or may not be achieved
(easier to learn to read and write), or some functional simplicity (can use the
standard ASCII symbol system on the internet), but there exists no concept of
‘truth’. We would not achieve any deeper ‘insight’ into the syntax and morphology
of the languages.
To summarize Dr. Tarnawsky’s criticisms in one sentence: Basically, all
proposed semantic representations (as presented by Katz, et. al.) are simply
systems that translate from one syntactic notation into another.
More complex mathematical examples give more complex analogies, like
those discussed by Putnam and his followers. Following Rene Descartes, x2 + y
2
=1, an algebraic expression in the domain of numbers, defines a circle in Cartesian
coordinates in the geometric domain of Euclid. Descartes showed that many
standard geometric shapes have a representation as polynomials. But this is simply
‘translation’ from one notation into another. These ‘translations’ simply relate
synonymous expressions and give no insight into what the reality underlying the
various notations might be. To understand the broader implications of these issues
examine the Church-Turing Hypothesis.
26
Dr. Tarnawsky, with his strong computational and electrical engineering
background, understood all the examples discussed by the Putnam group of
philosophers and mathematicians. Considering his background of NYU graduate
school coursework, research at IBM, and our many one on one personal discussion,
it became obvious to us both that he could write a dissertation discussing the
various views of the sound-meaning correspondences, do it at the full level of
mathematical/logical depth required, and further, offer a significant formal
representation of a Putnam-type semantics linked to a Chomsky generative
grammar syntax. He did a great job.
Following Putnam’s views Dr. Tarnawsky develops a systematic and
coherent perspective in which to present ‘meaning’ as a process that defines
‘correlations’ of sentences with complex systems that may not be ‘symbolic
systems’ like human language. For Dr. Tarnawsky semantic interpretation
involves more than pencil and paper ‘translations’ into other notations. Consider
two widely discussed ‘correlations’: one involving ‘action’, the other ‘social
integration of meanings’. Neither of these was discussed in Chomsky linguistic
circles. Dr. Tarnawsky’s work defined a ‘bridge’ to integrate them into the
Chomsky syntactic grammar perspective.
‘The blocks world is one of the most famous planning domains in artificial
intelligence. Imagine a set of cubes (blocks) sitting on a table. The goal is to build
one of more vertical stacks of blocks. The catch is that only one block may be
moved at a time: it may either be placed on the table or placed atop another block.
Because of this, any blocks that are at a given time under another block, cannot be
moved.
27
‘The simplicity of this toy world lends itself readily to symbolic or classical
AI approaches, in which the world is modeled as a set of abstract symbols which
may be reasoned about.’ (Wikipedia, Blocks world)
The researchers that developed the natural language interface to the blocks
world and the structure of the blocks world and movement therein were Terry
Winograd and Patrick Winston. Their initial studies appeared around 1970 and
developed in several places for the next ten+ years. If one typed in: ‘Put the red
block on top of the blue block’, the system would perform the simple movement,
or, if the blue block had something already on it, the program would remove the
object and then place the red block on top. If initially the red block laid under the
blue block various movements would occur. This type of semantic-pragmatic
process aligns well with the ideas of Putnam and Dr. Tarnawsky’s systematic
analysis. A sentence correlates with certain movements in the blocks world with
well-defined start and end states.
To consider one analysis that became quite popular: Suppose you are
wearing pants with pennies in the back right pocket and nickels in the front right
pocket. You must switch their positions. You will need two hands. Empty the right
rear pocket with the right hand and put the pennies into the left hand. Move the
nickels to the back pocket using the right hand. Then move the pennies from the
left hand to the right hand and into the front pocket. Of course, you could start with
the front pocket, so pragmatically at the level of robotic actions, the command to
‘switch the contents of the pockets’ has a functional ambiguity given the context.
Suppose the pennies are in the left front pocket and the nickels in the right
front pocket: You must switch them. This is difficult (impossible) unless there
28
exists a ‘storage location’ for one hand to place its contents so the hands can switch
coins from one to the other. The command: ‘Switch the pennies and nickels from
one pocket to the other’ differs radically in ‘computational complexity’ given the
physical situation. In the 1990’s some economists picked up on ‘blocks world’
thinking to analyze how to switch ‘wealth’ in dollars into various European and
Asian currencies, vice versa, and perhaps using gold as an intermediate storage
mechanism. A physical paper hundred dollar bill has no more intrinsic value than a
paper one dollar bill, and functions only as a symbol of wealth. Not so with gold,
which in its physical form, constituted wealth. The issues are somewhat like
converting from potential to kinetic energy. Putnam’s research reached into many
fields. Many of the ‘modern’ issues concerning ‘electronic money’, such as
bitcoins, were hashed out by Putnam’s students and colleagues decades ago.
Dr. Tarnawsky’s ‘social meaning’ approach lines up completely with
Putnam’s views on how meanings seem to be held by ‘groups of people’. Vast
amounts of articles, lectures, and books discuss this in many domains of inquiry.
Suppose one goes to the doctor with some painful ailment. The patient describes
the problem using simple English: ‘I suffer sharp pains in the big toe after eating
triple crème cheeses and washing them down with Merlot’. The doctor correlates
several thoughts into a diagnosis, cure, and prognosis. Probably gout
inflammations. (Indicative) Stop feasting on cheese and wine and start eating
salads and drink water! (Imperative) Are you on Medicare?; if so, you must take
these pills. (Ostensive: handing over of pills) Do you have a solid gold health
insurance?; if so, go to a spa in France. (Speculative future action). Another doctor
or nutritionist might give totally different responses. A religiously inclined advisor
(Pascal) might say ‘Welcome the pain. Suffering brings us closer to God.’ Notice
the patient’s simple indicative sentence has ‘meanings’ that are questions,
29
imperatives, ostensive definitions… and many more types of verbalizations and
actions.
Many people after visiting a doctor visit a pharmacist who gives them pills
with unpronounceable names and instructions on how many to eat, when, and with
or without food. Some go to X-ray facilities or other places with complex
machines they might have to climb into. All of the above concepts add up to give
the meaning of the patient’s claim about ‘sharp pains….’ Here we think of
Putnam’s discussion of elm versus beech trees. We have no clear ideas in our
heads about the objective reality of chemicals and machines that relate to our
pains. We only have the subjective knowledge of the pains, and of these, we can
rank them for intensity, type (sharp, dull…), and along other dimensions. To
reflect on more complex systems: What does it mean to yell ‘Fire!!!’? Why did
Chief Justice Oliver Wendell Holmes argue that ‘freedom of speech’ does not
include the right to yell ‘fire’ in a crowded room if there is no ‘physical fire’
present? Smoldering cigars and pipes, while burning substances, did not qualify as
‘fire’ for this utterance, and could not legitimize the exclamation.
Dr. Tarnawsky’s work appeared as ‘rebellious’ within the Chomsky
linguistics researchers at MIT, who all favored ‘interpretive semantics’ and
‘representations’, all of which involved translations between symbol systems.
Since Dr. Tarnawsky’s dissertation, much computational linguistic work has
followed in paths laid down by Putnam. Dr. Tarnawsky’s dissertation might well
have been the first work to integrate Putnam’s and Chomsky’s approaches into one
formulation. Work has continued in this direction. Let us look at one recent
development.
30
Nisan and Schocken’s The Elements of Computing Systems: Building a Modern
Computer From First Principles (2005) provides a modern example, albeit
somewhat conceptually difficult. In this brilliant book, they describe how one
can—within the system they describe—take an English sentence and convert it to a
logical notation. They then correlate this notation with transistor and electronic
circuitry to form very basic logical/electrical functional computational systems.
Then they analyze these somewhat complex functional units into the lowest level
of computational units, NAND gates. Then, from the most elemental NAND gate,
they show how to construct a complete modern computer (a Turing machine) that
has an input and an output and can solve problems of all sorts. Nisan and
Schocken’s book represents one of the finest conceptual schemes to arise from the
types of thinking in Putnam’s philosophy and in the semantic action and social
systems discussed by Dr. Tarnawsky.
Ray C. Dougherty
Professor Emeritus
Department of Linguistics
New York University